Neuro-Symbolic Traders: Assessing the Wisdom of AI Crowds in Markets

TL;DR

This work develops a form of virtual traders that use deep generative models to make buy/sell decisions, which are term neuro-symbolic traders, and exposes several groups of neuro-symbolic traders to a virtual market environment.

Abstract

Deep generative models are becoming increasingly used as tools for financial analysis. However, it is unclear how these models will influence financial markets, especially when they infer financial value in a semi-autonomous way. In this work, we explore the interplay between deep generative models and market dynamics. We develop a form of virtual traders that use deep generative models to make buy/sell decisions, which we term neuro-symbolic traders, and expose them to a virtual market. Under our framework, neuro-symbolic traders are agents that use vision-language models to discover a model of the fundamental value of an asset. Agents develop this model as a stochastic differential equation, calibrated to market data using gradient descent. We test our neuro-symbolic traders on both synthetic data and real financial time series, including an equity stock, commodity, and a foreign exchange pair. We then expose several groups of neuro-symbolic traders to a virtual market environment. This market environment allows for feedback between the traders belief of the underlying value to the observed price dynamics. We find that this leads to price suppression compared to the historical data, highlighting a future risk to market stability. Our work is a first step towards quantifying the effect of deep generative agents on markets dynamics and sets out some of the potential risks and benefits of this approach in the future.

Figures (4)

• Figure 1: Schematic for our neuro-symbolic traders that are exposed to a virtual market. A neuro-symbolic agent consists of a visual-language model (VLM) which can output model code for a stochastic differential equation (SDE) based on some observed price path image and instruction (A). The model is fit to the data using either a gradient-based or VLM-based calibration (B). Model output and observations are then passed to a second VLM which evaluates the model performance and suggests improvements (C). This is then passed to the original VLM and the loop continues (D). The steps within the virtual market are outlined in the figure.
• Figure 2: Comparison of different methods used to fit a stochastic differential equation to time series data, including (a) a vision-language model (VLM) approach where parameter values are suggested based on the image and model description, (b) a gradient-based approach where the error is directly minimised using differentiable programming, and (c-d) which introduces additional instructions to the VLM to construct simple or parsimonious models.
• Figure 3: Comparison of models for the fundamental value proposed by a neuro-symbolic trader for three different asset classes and comparing both with and without domain information passed to the trader.
• Figure 4: Simulated price path (dotted line) generated by neuro-symbolic traders model of the fundamental value (faded lines) within a virtual market. Each section, marked by a different colour and separated by vertical lines, relates to periods used to fit the model ($t_{i-1}$) and then simulate the impact in price ($t_i$). The original S&P 500 stock is included for reference (solid line).